LSJ1309 - Facilitating Effective User Navigation Through Website Structure Improvement

Facilitating Effective User Navigation through
Website Structure Improvement

Min Chen and Young U. Ryu
AbstractDesigning well-structured websites to facilitate effective user navigation has long been a challenge. A primary reason is that
the web developers understanding of how a website should be structured can be considerably different from that of the users. While
various methods have been proposed to relink webpages to improve navigability using user navigation data, the completely
reorganized new structure can be highly unpredictable, and the cost of disorienting users after the changes remains unanalyzed. This
paper addresses how to improve a website without introducing substantial changes. Specifically, we propose a mathematical
programming model to improve the user navigation on a website while minimizing alterations to its current structure. Results from
extensive tests conducted on a publicly available real data set indicate that our model not only significantly improves the user
navigation with very few changes, but also can be effectively solved. We have also tested the model on large synthetic data sets to
demonstrate that it scales up very well. In addition, we define two evaluation metrics and use them to assess the performance of the
improved website using the real data set. Evaluation results confirm that the user navigation on the improved structure is indeed greatly
enhanced. More interestingly, we find that heavily disoriented users are more likely to benefit from the improved structure than the less
disoriented users.
Index TermsWebsite design, user navigation, web mining, mathematical programming
1 INTRODUCTION
T
HE advent of the Internet has provided an unprecedented
platform for people to acquire knowledge and explore
information. There are 1.73 billion Internet users worldwide
as of September 2009, an increase of 18 percent since 2008 [1].
The fast-growing number of Internet users also presents
huge business opportunities to firms. According to Grau [2],
the US retail e-commerce sales (excluding travel) totaled
$127.7 billion in 2007 and will reach $218.4 billion by 2012. In
order to satisfy the increasing demands from online
customers, firms are heavily investing in the development
and maintenance of their websites. InternetRetailer [3]
reports that the overall website operations spending
increased in 2007, with one-third of site operators hiking
spending by at least 11 percent, compared to that in 2006.
Despite the heavy and increasing investments in website
design, it is still revealed, however, that finding desired
information in a website is not easy [4] and designing
effective websites is not a trivial task [5], [6]. Galletta et al.
[7] indicate that online sales lag far behind those of brick-
and-mortar stores and at least part of the gap might be
explained by a major difficulty users encounter when
browsing online stores. Palmer [8] highlights that poor
website design has been a key element in a number of high
profile site failures. McKinney et al. [9] also find that users
having difficulty in locating the targets are very likely to
leave a website even if its information is of high quality.
A primary cause of poor website design is that the web
developers understanding of how a website should be
structured can be considerably different from those of the
users [10], [11]. Such differences result in cases where users
cannot easily locate the desired information in a website.
This problem is difficult to avoid because when creating a
website, web developers may not have a clear under-
standing of users preferences and can only organize pages
based on their own judgments. However, the measure of
website effectiveness should be the satisfaction of the users
rather than that of the developers. Thus, Webpages should
be organized in a way that generally matches the users
model of how pages should be organized [12].
Previous studies on website has focused on a variety of
issues, such as understanding web structures [13], finding
relevant pages of a given page [14], mining informative
structure of a news website [15], [16], and extracting
template from webpages [17]. Our work, on the other hand,
is closely related to the literature that examines how to
improve website navigability through the use of user naviga-
tion data. Various works have made an effort to address this
question and they can be generally classified into two
categories [11]: to facilitate a particular user by dynamically
reconstituting pages based on his profile and traversal
paths, often referred as personalization, and to modify the
site structure to ease the navigation for all users, often
referred as transformation.
In this paper, we are concerned primarily with transfor-
mation approaches. The literature considering transforma-
tions approaches mainly focuses on developing methods to
completely reorganize the link structure of a website.
Although there are advocates for website reorganization
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 25, NO. 3, MARCH 2013 571
. M. Chen is with the School of Management, George Mason University,
Fairfax, VA 22030. E-mail: mchen15@gmu.edu.
. Y.U. Ryu is with the Naveen Jindal School of Management, The University
of Texas at Dallas, SM33, 800 West Campbell Road, Richardson, Texas
75080-3021. E-mail: ryoung@utdallas.edu.
Manuscript received 18 Nov. 2010; revised 12 Oct. 2011; accepted 23 Oct.
2011; published online 16 Nov. 2011.
For information on obtaining reprints of this article, please send e-mail to:
tkde@computer.org, and reference IEEECS Log Number TKDE-2010-11-0606.
Digital Object Identifier no. 10.1109/TKDE.2011.238.
1041-4347/13/$31.00 2013 IEEE Published by the IEEE Computer Society
approaches, their drawbacks are obvious. First, since a
complete reorganization could radically change the location
of familiar items, the new website may disorient users [18].
Second, the reorganized website structure is highly un-
predictable, and the cost of disorienting users after the
changes remains unanalyzed. This is because a websites
structure is typically designed by experts and bears business
or organizational logic, but this logic may no longer exist in
the new structure when the website is completely reorga-
nized. Besides, no prior studies have assessed the usability
of a completely reorganized website, leading to doubts on
the applicability of the reorganization approaches. Finally,
since website reorganization approaches could dramatically
change the current structure, they cannot be frequently
performed to improve the navigability.
Recognizing the drawbacks of website reorganization
approaches, we address the question of how to improve the
structure of a website rather than reorganize it substantially.
Specifically, we develop a mathematical programming (MP)
model that facilitates user navigation on a website with
minimal changes to its current structure. Our model is
particularly appropriate for informational websites whose
contents are static and relatively stable over time. Examples
of organizations that have informational websites are
universities, tourist attractions, hospitals, federal agencies,
and sports organizations. Our model, however, may not be
appropriate for websites that purely use dynamic pages or
have volatile contents. This is because a steady state might
never be reached in user access patterns in such websites, so
it may not be possible to use the weblog data to improve the
site structure [19].
The number of outward links in a page, i.e., the out-
degree, is an important factor in modeling web structure.
Prior studies typically model it as hard constraints so that
pages in the new structure cannot have more links than a
specified out-degree threshold, because having too many
links in a page can cause information overload to users and
is considered undesirable. For instance, Lin [20] uses 6, 8,
and 10 as the out-degree threshold in experiments. This
modeling approach, however, enforces severe restrictions
on the new structure, as it prohibits pages from having
more links than a specified threshold, even if adding these
links may greatly facilitate user navigation. Our model
formulates the out-degree as a cost term in the objective
function to penalize pages that have more links than the
threshold, so a pages out-degree may exceed the threshold
if the cost of adding such links can be justified.
We perform extensive experiments on a data set
collected from a real website. The results indicate that our
model can significantly improve the site structure with only
few changes. Besides, the optimal solutions of the MP
model are effectively obtained, suggesting that our model is
practical to real-world websites. We also test our model
with synthetic data sets that are considerably larger than the
real data set and other data sets tested in previous studies
addressing website reorganization problem. The solution
times are remarkably low for all cases tested, ranging from a
fraction of second to up to 34 seconds. Moreover, the
solution times are shown to increase reasonably with the
size of the website, indicating that the proposed MP model
can be easily scaled to a large extent.
To assess the user navigation on the improved website,
we partition the entire real data set into training and testing
sets. We use the training data to generate improved
structures which are evaluated on the testing data using
simulations to approximate the real usage. We define two
metrics and use them to assess whether user navigation is
indeed enhanced on the improved structure. Particularly,
the first metric measures whether the average user naviga-
tion is facilitated in the improved website, and the second
metric measures how many users can benefit from the
improved structure. Evaluation results confirm that user
navigation on the improved website is greatly enhanced.
In summary, this paper makes the following contribu-
tions. First, we explore the problem of improving user
navigation on a website with minimal changes to the
current structure, an important question that has never
been examined in the literature. We show that our MP
model not only successfully accomplishes the task but also
generates the optimal solutions surprisingly fast. The
experiments on synthetic data indicate that our model also
scales up very well. Second, we model the out-degree as a
cost term in the objective function instead of as hard
constraints. This allows a page to have more links than the
out-degree threshold if the cost is reasonable and hence
offers a good balance between minimizing changes to a
website and reducing information overload to users. Third,
we propose two evaluation metrics and use them to assess
the improved structure to confirm the validity of our
model. The evaluation procedure developed in this paper
provides a framework for evaluating website structures in
similar studies.
The rest of the paper is organized as follows: Section 2
reviews related literature. Section 3 introduces the metric
for evaluating user navigation and describes the problem.
Section 4 presents the model formulation with several
illustrative examples. Section 5 describes the data set,
analyzes the results of computational experiments, and
reports the evaluation results. Section 6 discusses issues
related to this research, and Section 7 concludes the paper.
2 RELATED WORK
The growth of the Internet has led to numerous studies on
improving user navigations with the knowledge mined
from webserver logs and they can be generally categorized
in to web personalization and web transformation
approaches [11].
Web personalization is the process of tailoring
webpages to the needs of specific users using the
information of the users navigational behavior and profile
data [21]. Perkowitz and Etzioni [11] describe an approach
that automatically synthesizes index pages which contain
links to pages pertaining to particular topics based on the
co-occurrence frequency of pages in user traversals, to
facilitate user navigation. The methods proposed by
Mobasher et al. [22], [23], [24] and Yan et al. [25] create
clusters of users profiles from weblogs and then dynami-
cally generate links for users who are classified into
different categories based on their access patterns.
572 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 25, NO. 3, MARCH 2013
Nakagawa and Mobasher [26] develop a hybrid persona-
lization system that can dynamically switch between
recommendation models based on degree of connectivity
and the users position in the site. For reviews on web
personalization approaches, see [21] and [27].
Web transformation, on the other hand, involves chan-
ging the structure of a website to facilitate the navigation for
a large set of users [28] instead of personalizing pages for
individual users. Fu et al. [29] describe an approach to
reorganize webpages so as to provide users with their
desired information in fewer clicks. However, this approach
considers only local structures in a website rather than the
site as a whole, so the new structure may not be necessarily
optimal. Gupta et al. [19] propose a heuristic method based
on simulated annealing to relink webpages to improve
navigability. This method makes use of the aggregate user
preference data and can be used to improve the link
structure in websites for both wired and wireless devices.
However, this approach does not yield optimal solutions
and takes relatively a long time (10 to 15 hours) to run even
for a small website. Lin [20] develops integer programming
models to reorganize a website based on the cohesion
between pages to reduce information overload and search
depth for users. In addition, a two-stage heuristic involving
two integer-programming models is developed to reduce
the computation time. However, this heuristic still requires
very long computation times to solve for the optimal
solution, especially when the website contains many links.
Besides, the models were tested on randomly generated
websites only, so its applicability on real websites remains
questionable. To resolve the efficiency problem in [20], Lin
and Tseng [28] propose an ant colony system to reorganize
website structures. Although their approach is shown to
provide solutions in a relatively short computation time, the
sizes of the synthetic websites and real website tested in [28]
are still relatively small, posing questions on its scalability
to large-sized websites.
There are several remarkable differences between web
transformation and personalization approaches. First,
transformation approaches create or modify the structure
of a website used for all users, while personalization
approaches dynamically reconstitute pages for individual
users. Hence, there is no predefined/built-in web structure
for personalization approaches. Second, in order to under-
stand the preference of individual users, personalization
approaches need to collect information associated with
these users (known as user profiles). This computationally
intensive and time-consuming process is not required for
transformation approaches. Third, transformation ap-
proaches make use of aggregate usage data from weblog
files and do not require tracking the past usage for each user
while dynamic pages are typically generated based on the
users traversal path. Thus, personalization approaches are
more suitable for dynamic websites whose contents
are more volatile and transformation approaches are more
appropriate for websites that have a built-in structure and
store relatively static and stable contents.
This paper examines the questions of how to improve
user navigation in a website with minimal changes to its
structure. It complements the literature of transformation
approaches that focus on reconstructing the link structure of
a website. As a result, our model is suitable for website
maintenance and can be applied in a regular manner.
3 METRIC FOR EVALUATING NAVIGATION
EFFECTIVENESS
3.1 The Metric
Our objective is to improve the navigation effectiveness of a
website with minimal changes. Therefore, the first question
is, given a website, how to evaluate its navigation effective-
ness. Marsico and Levialdi [30] point out that information
becomes useful only when it is presented in a way
consistent with the target users expectation. Palmer [31]
indicates that an easy-navigated website should allow users
to access desired data without getting lost or having to
backtrack. We follow these ideas and evaluate a websites
navigation effectiveness based on how consistently the
information is organized with respect to the users expectations.
Thus, a well-structured website should be organized in
such a way that the discrepancy between its structure and
users expectation of the structure is minimized. Since users
of informational websites typically have some information
targets [19], [32], i.e., some specific information they are
seeking, we measure this discrepancy by the number of
times a user has attempted before locating the target.
Our metric is related to the notion of information scent
developed in the context of information foraging theory [33],
[34], [35]. Information foraging theory models the cost
structure of human information gathering using the
analogy of animals foraging for food and is a widely
accepted theory for addressing the information seeking
process on the web [32], [36], [37], [38], [39]. Information
scent refers to proximal cues (e.g., the snippets of text and
graphics of links) that allow users to estimate the location
of the distal target information and determine an
appropriate path [34]. Users are faced with a decision
point at each page; they use information scent to evaluate
the likely effort and the probability of reaching their
targets via each link and make navigation decisions
accordingly [7]. Consequently, a user is assumed to follow
the path that appears most likely to lead him to the target.
This suggests that a user may backtrack to an already
visited page to traverse a new path if he could not locate
the target page in the current path. Therefore, we use the
number of paths a user has traversed to reach the target as a
proximate measure to the number of times the user has
attempted to locate one target.
We use backtracks to identify the paths that a user has
traversed, where a backtrack is defined as a users revisit to a
previously browsed page. The intuition is that users will
backtrack if they do not find the page where they expect it
[40]. Thus, a path is defined as a sequence of pages visited
by a user without backtracking, a concept that is similar to
the maximal forward reference defined in Chen et al. [41].
Essentially, each backtracking point is the end of a path.
Hence, the more paths a user has traversed to reach the
target, the more discrepant the site structure is from the
users expectation.
CHEN AND RYU: FACILITATING EFFECTIVE USER NAVIGATION THROUGH WEBSITE STRUCTURE IMPROVEMENT 573
3.2 An Example
We use an example to illustrate the aforementioned
concepts and how to extract the metric from weblog files.
To analyze the interaction between users and a website, the
log files must be broken up into user sessions. Cooley et al.
[42] define a session as a group of activities performed by a
user during his visit to a site and propose timeout methods
to demarcate sessions from raw log files. In this definition, a
session may include one or more target pages, as a user may
visit several targets during a single session. Since the metric
used in our analysis is the number of paths traversed to find
one target, we use a different term mini session to refer to a
group of pages visited by a user for only one target. Thus, a
session may contain one or more mini sessions, each of
which comprises a set of paths traversed to reach the target.
We use the page-stay timeout heuristic described in [40],
[43] to demarcate mini sessions. Specifically, we identify
whether a page is the target page by evaluating if the time
spent on that page is greater than a timeout threshold. The
intuition is that a user generally spends more time reading
on the documents that they find relevant than those they do
not [44]. Though it is impossible to identify user sessions
unerringly from weblog files [19], we find the page-stay
heuristic an appropriate technique for the context of our
problem and we provide a detailed discussion on this
heuristic in Section 6.
We depict in Fig. 1 a hypothetical website that has
10 pages. Fig. 2 illustrates a mini session, where a user
starts from A, browses D and H, and backtracks to D, from
where he visits C, B, E, J, and backtracks to B. Then, this
user goes from B to F and finally reaches the target K. We
formally denote the mini session by S = A; D; H;
C; B; E; J; F; K, where an element in S represents a
path traversed by the user. In this example, mini session S
has three paths as the user backtracks at H and J before
reaching the target K. Note that D and B only appear once
in S because of caching.
3.3 Problem Description
Difficulty in navigation is reported as the problem that
triggers most consumers to abandon a website and switch
to a competitor [45]. Generally, having traversed several
paths to locate a target indicates that this user is likely to
have experienced navigation difficulty. Therefore, Web-
masters can ensure effective user navigation by improving
the site structure to help users reach targets faster. This is
especially vital to commercial websites, because easy-
navigated websites can create a positive attitude toward
the firm, and stimulate online purchases [46], whereas
websites with low usability are unlikely to attract and
retain customers [47].
Our model allows Webmasters to specify a goal for user
navigation that the improved structure should meet. This
goal is associated with individual target pages and is
defined as the maximum number of paths allowed to reach the
target page in a mini session. We term this goal the path
threshold for short in this paper. In other words, in order to
achieve the user navigation goal, the website structure must
be altered in a way such that the number of paths needed to
locate the targets in the improved structure is not larger
than the path threshold.
In the example shown in Fig. 2, the user has traversed
three paths before reaching the target. An intuitive solution
to help this user reach the target faster is to introduce more
links [10], [19], [48]. There are many ways to add extra links.
If a link is added from D to K, the user can directly reach K
via D, and hence reach the target in the first path. Thus,
adding this link saves the user two paths. Similarly,
establishing a link from B to K enables the user to reach the
target in the second path. Hence, this saves him one path.
We could also insert a link from E to K, and this is
considered the same as linking B to K. This is because both
B and E are pages visited in the second path, so linking
either one to K saves only one path. Yet, another possibility
is to link C to F, a nontarget page. In this case, we assume
that the user does not follow the new link, because it does
not directly connect a page to the target.
While many links can be added to improve navigability,
our objective is to achieve the specified goal for user
navigation with minimal changes to a website. We measure
the changes by the number of new links added to the
current site structure. There are several reasons that we
should insert minimal links. First, minimizing changes to
the current structure can avoid disorienting familiar users.
Second, adding unnecessary links can lead to pages having
too many links, which increases users cognitive loads and
makes it difficult for them to read and comprehend [49].
Third, since our model improves site structures on a regular
basis, the number of new links should be kept at minimum
Fig. 1. A website with 10 pages.
Fig. 2. Example of a mini session.
such that the links in a website in the whole course of
maintenance do not expand in a chaotic manner.
There are cases where users could have reached the
targets through existing links, but failed to do so in
practice. One reason could be that these links are placed
in inconspicuous locations and hence are not easily notice-
able. Another reason might be that the labels of these links
are misleading or confusing, causing difficulty to users in
predicting the content at the target page [50]. As a result,
Webmasters should focus on enhancing the design of these
existing links before adding new links. Our model
considers this issue by placing a preference on the selection
of such existing links.
4 PROBLEM FORMULATION
4.1 The Model
Our problem can be regarded as a special graph
optimization problem. We model a website as a directed
graph, with nodes representing pages and arcs represent-
ing links. Let N be the set of all webpages and
ij
, where
i; j N, denote page connectivity in the current structure,
with
ij
= 1 indicating page i has a link to page j, and
ij
= 0 otherwise. The current out-degree for page i is
denoted by W
i
=
P
jN

ij
.
From the log files, we obtain the set T of all mini
sessions. For a mini session S T, we denote tgt(S) the
target page of S. Let L
m
(S) be the length of S, i.e., the
number of paths in S, and L
p
(k; S), for 1 _ k _ L
m
(S),
be the length of the kth path in S, i.e., the number of pages
in the kth path of S. We further define docno(r; k; S), for
1 _ k _ L
m
(S) and 1 _ r _ L
p
(k; S), as the rth page visited
in the kth path in S. Take the mini session S in Fig. 2 for
example, it follows that L
m
(S) = 3; L
p
(1; S) = 3, and
docno(1; 1; S) = A, as this mini session has three paths
and the first path has three pages (A, D, and H) in which
page A is the first page. We define E = (i; j) : i; j N and
S T such that i S and j = tgt (S) and N
E
= i : (i;
j) E. In essence, E is the set of candidate links that can be
selected to improve the site structure to help users reach
their targets faster. Our problem is to determine whether to
establish a link from i to j for (i; j) E. Let x
ij
0; 1
denote the decision variable such that x
ij
= 1 indicates
establishing the link.
As explained earlier, Webmasters can set a goal for user
navigation for each target page, which is denoted by b
j
and
is termed the path threshold for page j. Given a mini session
S with target page j and a path threshold b
j
, we can
determine whether the user navigation goal is achieved in S
by comparing the length of S, i.e., L
m
(S), with path
threshold (b
j
) for the target page of S. If the length of S is
larger than b
j
, it indicates the user navigation in S is
below the goal. Then, we need to alter the site structure to
improve the user navigation in S to meet the goal.
Otherwise, no improvement is needed for S.
Intuitively, given path thresholds, we can determine
which mini sessions need to be improved and hence are
relevant to our decision (termed relevant mini sessions).
The irrelevant mini session are not considered in our
model. We denote the set of relevant mini sessions by
T
R
_ T. For a mini session S T
R
, it is said to be improved
if the website is altered in a way such that the user could
reach the target within the associated path threshold after
changes are made to the site structure.
We define parameters a
S
ijkr
to be 1 if docno(r; k; S) = i
and tgt(S) = j, and 0 otherwise. In other words, a
S
ijkr
= 1 if
and only if page i is the rth visited page in the kth path in
S and page j is the target of S. Further, we define variable
c
S
kr
which will be set to one if the solution indicates
establishing a link from the rth page in the kth path in S to
the target page of S, i.e., tgt(S), and 0 otherwise. As will be
explained later, the use of a
S
ijkr
is to build connections
between variables x
ij
and c
S
kr
, where the first variable uses
global indices and the latter is defined using local indices.
Similar to prior studies, appropriate out-degree thresh-
olds can be specified for webpages. We denote C
i
the out-
degree threshold for page i. Nevertheless, our model
penalizes a page if its out-degree is larger than the
threshold instead of modeling the threshold as a hard
constraint. In effect, out-degree C
i
indicates the maximum
number of links that page i can have without being
penalized. Let p
i
be the number of links exceeding the
out-degree threshold C
i
for page i in the improved
structure. Depending on the application of our model,
different weights of penalties can be imposed on pages
whose out-degree exceeds the respective out-degree thresh-
old. We denote the weight by m and term it the multiplier for
the penalty term. Table 1 provides a summary of the
notations used in this paper.
The problem of improving the user navigation on a
website while minimizing the changes to its current
structure can then be formulated as the mathematical
programming model below:
Minimize
X
(i;j)E
x
ij
[1
ij
(1 ")[ m
X
iN
E
p
i
subject to
c
S
kr
=
X
(i;j)E
a
S
ijkr
x
ij
; r = 1; 2; . . . ; L
p
(k; S);
k = 1; 2; . . . ; L
m
(S); \S T
R
(1)
X
b
j
k=1
X
L
p
(k;S)
r=1
c
S
kr
_ 1; \S T
R
; j = tgt(S) (2)
X
j:(i;j)E
x
ij
(1
ij
) W
i
p
i
_ C
i
; \i N
E
(3)
x
ij
0; 1; p
i
0 Z
; \(i; j) E; i N
E
: (4)
The objective function minimizes the cost needed to
improve the website structure, where the cost consists of
two components: 1) the number of new links to be
established (the first summation), and 2) the penalties on
pages containing excessive links, i.e., more links than the
out-degree threshold (C
i
), in the improved structure
(the second summation).
We have noted that some existing links may often be
neglected by users due to poor design or ambiguous labels.
Such links should be improved first before any new links
are established. Therefore, we introduce [1
ij
(1 ")[,
where " is a very small number, in the objective function to
let the model select existing links whenever possible. Note
that if (1 ") is not present, then there is no cost in choosing
an existing link, and this could lead to a number of optima.
As an extreme example, if (1 ") is removed and the
penalty term is not included, the costs of establishing new
links, i.e.,
P
(i;j)E
x
ij
(1
ij
) when selecting all existing
links are the same as the costs when none of them is
selected. This occurs because there is no cost in selecting an
existing link, i.e., (1
ij
) = 0, when
ij
= 1. Thus, we add
(1 ") to impose a very small cost on improving an existing
link such that the model will select the minimal number of
existing links for improvement.
Constraint (1) defines variable c
S
kr
, which is set to 1 if and
only if a
S
ijkr
= 1 and x
ij
= 1, for some (i; j) E, and 0
otherwise. It uses parameter a
S
ijkr
to build connections
between variables x
ij
and c
S
kr
, because the objective function
is defined by variable x
ij
which uses global indices (i and j)
to label webpages, whereas the constraint (2) is defined by
variable c
S
kr
which uses local indices (S, k, and r) to identify
a pages position in a mini session. The values of a
S
ijkr
can be
easily obtained after mini sessions are identified from
weblog files, as demonstrated in an example later.
Constraint (2) requires that the goal for user navigation
be achieved for all relevant mini sessions, where the goal is
defined as path threshold (b
j
). Particularly, for a mini
session S T
R
in which the user navigation is below the
specified goal, i.e., L
m
(S) > b
j
for j = tgt(S), at least one link
from pages visited on or before the b
j
th path to the target j is
either established or improved so that the user can reach the
page j within the path threshold set by the Webmaster.
Constraint (3) uses p
i
to capture the number of links
exceeding the out-degree threshold C
i
for page i N
E
.
This value (p
i
) is then used to compute penalties in the
objective function. The degree of penalty can be controlled
by the multiplier for the penalty term (m). Constraint (4)
imposes that decision variables are binary and p
i
are
nonnegative integers.
In practice, parameter values such as b
j
and C
i
are
context dependent and can vary across webpages. How-
ever, for simplicity, we used a single value denoted by b as
the path threshold and denoted by C as the out-degree
threshold for all pages in our examples and experiments.
We discuss how to select appropriate parameter values in
detail in Section 6.
Note that a special case of our MP model when m = 0
and
ij
= 0 for i; j N can be viewed as the hitting set
problem. That is, when pages are not penalized for having
too many links and no preference is placed on selecting
existing links, our formulation reduces to a hitting set
problem. The hitting set problem is stated as follows: Given
a ground set X and a collection of subset F, the objective is
to find the smallest subset H _ X of elements that hits
every set of F, i.e., H A ,= O for every A F. It is
equivalent to the set-covering problem which is known to
be NP-complete [51].
In the context of our problem, E is the ground set
containing all candidate links that can be used for
improving navigability. For a relevant mini session S
T
R
with path threshold b
j
, denote the set of candidate links
that can be selected to help achieve the user navigation goal
in S by B = (i; j) : (i; j) E and a
S
ijkr
= 1, for j = tgt(S);
1 _ k _ b
j
, and 1 _ r _ L
p
(k; S). Further, we denote the
collection of the sets of such candidate links for all mini
sessions in T
R
by D. The objective of our formulation when
m = 0 and
ij
= 0 is to find the smallest subset H _ E of
elements that hits every set of D, i.e., H B ,= O for every
B D.
4.2 Observations on Reduction of Problem Size
The formulation has [E[ binary variables corresponding to
the number of candidate links and [T
R
[ constraints
TABLE 1
Summary of Notations
corresponding to the number of relevant mini sessions.
While in practice the size of a website and the number of
mini sessions obtained from server logs can be very large, it
turns out that, in the context of our problem, the
formulation can be reduced to a significantly smaller one
that can be quickly solved. We make several observations
related to the problem size. These observations together
provide insights into why the problem size in our
formulation can be considerably reduced and help explain
the fast solution times in our experiments in the later
sections. In fact, as will be shown later, our formulation has
already taken steps to reduce the problem size.
4.2.1 Relevant Mini Sessions
Recall that a mini session is relevant only if its length is
larger than the corresponding path threshold. Conse-
quently, only relevant mini sessions need to be considered
for improvement and this leads to a large number of
irrelevant mini sessions (denoted as T
I
) being eliminated
from consideration in our MP model. In other words, define
T
I
= TT
R
, any mini session S T
I
will not be considered
in our formulation as the user navigation in S already meets
the goal (set as path threshold).
As will be shown later, the choice of path threshold can
have significant impacts on relevant mini sessions.
Generally, increasing the path threshold leads to a smaller
number of relevant mini sessions while decreasing it has
the opposite effect. For example, for the real data set used
in the experiments (details are provided in Section 5.1),
when the path threshold (b) increases from 3 to 5, the
number of relevant mini sessions reduces from several
thousand to only a few hundred. Even for the case when
b = 1, a large number of irrelevant mini sessions can be
eliminated from consideration.
4.2.2 Relevant Candidate Links
Define E
/
= (i; j) : i; j N as the possible links between
all pages in a website with node set N. Theoretically, any
link from E
/
can be considered in our decision problem
without a preprocessing step, leading to a total number of
[N[ [N[ links (variables). This number can be very large
even for a small website. Intuitively, not every link in E
/
can
be used to improve user navigation. Recall that we term the
links that can be selected to help user navigate as candidate
links (denoted by E), which can be easily obtained from
mini sessions. Thus, it follows that x
ij
= 0 in the optimal
solution \(i; j) E
/
E. In other words, noncandidate links
do not help improve user navigation and hence need not
enter the formulation.
It turns out that many candidate links can also be
eliminated from consideration because they are not relevant
to the decision for two reasons. First, given path thresholds,
denote the set of candidate links for relevant mini sessions
by E
RM
and the set of candidate links for irrelevant mini
sessions by E
IM
. It follows that the candidate links in
E
IM
E
RM
are only for irrelevant mini sessions that need no
improvement and hence can be eliminated from considera-
tion. Second, not all candidate links in E
RM
might be
relevant to the decision. Particularly, for a mini session S
T
R
with path threshold b, a link is said to be relevant to S if
adding/improving it can help the user in S reach the target
in no more than b paths, i.e., achieve the user navigation goal
in S. Thus, the candidate links relevant to the decision are
those originating from the pages visited on the bth path or
before. The other candidate links for S can be eliminated
from consideration because selecting them cannot improve
S to achieve the specified goal for user navigation. We term
the set of candidate links relevant to the decision the relevant
candidate links, and we denote them by E
R
= (i; j) : (i; j)
E and S T
R
such that i S, j = tgt(S); a
S
ijkr
= 1 for 1 _
k _ b
j
; and 1 _ r _ L
p
(k; S). This leads to x
ij
= 0 in the
optimal solution \(i; j) E
RM
E
R
. Thus, the cardinality of
the set of relevant candidate links ([E
R
[) could be relatively
small even for a large website.
4.2.3 Dominated Mini Sessions
Another reason for the problem size reduction is that many
relevant mini sessions dominate others with respect to
relevant candidate links. Mini session S
p
dominates mini
session S
q
if the set of relevant candidate links for S
q
contains (at least) all relevant candidate links for S
p
. This
dominance is strict if there exists at least one candidate link
that is relevant to S
q
but is irrelevant to S
p
. Therefore, when
a mini session is improved in the new structure, the mini
sessions that are dominated by this one are also improved.
Consequently, the constraints corresponding to dominated
mini sessions are redundant and can be eliminated from
consideration in the MP model.
4.3 Illustrative Examples
In this section, we use examples to illustrate our model and
how the choices of different parameter values could affect
the problem size of the model, the solution spaces, and the
solutions. Let the matrix in Fig. 3 represent the current
connectivity of a website that has six pages. An entry value
indicates if a row node has a link to a column node, with 1
indicating a link exists. The out-degree of page i can be
obtained by summing the entry values in row n
i
. For
example, the matrix shows page 1 has a link to pages 4 and
5, i.e.,
14
=
15
= 1, and hence its out-degree is W
1
= 2.
Table 2 shows a set T of six mini sessions, each of which
has two or more paths. For example, mini session S
1
has
three paths and the target is page 6. The user starts from
page 2 and backtracks twice at pages 1 and 4, respectively,
because neither page has a link to 6. In the third path,
the user finally locates the target after visiting page 5. We
provide in Table 3 the set of candidate links that can be
added or improved to help users reach the targets faster for
each of the six mini sessions.
Now, consider a case where the Webmaster sets the path
threshold to b = 1, which means that the site structure
Fig. 3. The Connectivity matrix for illustrative examples.
should be improved such that users can reach their targets in
one path. In this case, all six mini session are relevant as their
lengths are all larger than the path threshold. This requires
that for each mini session, at least one link from the pages
visited in the first path to the target be established or
improved. In other words, only the candidate links originat-
ing from the pages visited in the first path are relevant to our
decision. Table 4 lists the set of relevant candidate links for
relevant mini sessions.
For the purpose of illustration, we do not consider the
penalty term for now. The problem is formulated as
Minimize
X
(i;j)E
x
ij
[1
ij
(1 ")[
subject to
c
S
kr
=
X
(i;j)E
a
S
ijkr
x
ij
; r = 1; 2; . . . ; L
p
(k; S);
k = 1; 2; . . . ; L
m
(S); \S T
R
(1)
c
S
1
11
c
S
1
12
_ 1 (S1)
c
S
2
11
c
S
2
12
_ 1 (S2)
c
S
3
11
c
S
3
12
c
S
3
13
_ 1 (S3)
c
S
4
11
c
S
4
12
_ 1 (S4)
c
S
5
11
c
S
5
12
_ 1 (S5)
c
S
6
11
c
S
6
12
c
S
6
13
_ 1 (S6)
x
ij
0; 1; \(i; j) E:
The first constraint connects x
ij
with c
S
kr
, such that when
x
ij
= 1 and a
S
ijkr
= 1, the corresponding c
S
kr
is also set to 1.
The a
S
ijkr
values can be easily obtained after mini sessions
are extracted from log data. For example, it follows from
Table 2 that a
S
1
2611
= 1, because the first page in the first path
of S
1
is page 2 and the target is page 6. As a result, a
solution that links pages 2 to 6, i.e., x
26
= 1, will also set
c
S
1
11
= 1, and hence satisfy constraint (S
1
). The other values
of a
S
ijkr
can be obtained similarly. We set " = 1:0E 8 in the
example. Solving this math program results in the optimal
solution x
26
= x
46
= x
14
= x
64
= 1, with the other variables
being 0. The only new link needed is (4, 6), with the others
being existing links, i.e.,
26
=
14
=
64
= 1.
The solution will change accordingly when we incorpo-
rate the penalty term into the objective function and change
the path threshold. For instance, if the out-degree threshold
and the multiplier for the penalty term are set as C = 3 and
m = 5, respectively, and the path threshold increases to
b = 2, then the new problem is formulated as follows:
Minimize
X
(i;j)E
x
ij
[1
ij
(1 ")[ 5
X
iN
E
p
i
subject to
c
S
kr
=
X
(i;j)E
a
S
ijkr
x
ij
; r = 1; 2; . . . ; L
p
(k; S);
k = 1; 2; . . . ; L
m
(S); \S T
R
(1)
c
S
1
11
c
S
1
12
c
S
1
21
_ 1 (S1)
c
S
2
11
c
S
2
12
c
S
2
21
c
S
2
22
_ 1 (S2)
c
S
4
11
c
S
4
12
c
S
4
21
c
S
4
22
_ 1 (S4)
c
S
5
11
c
S
5
12
c
S
5
21
_ 1 (S5)
X
j=1
x
ij
(1
ij
) W
i
p
i
_ 3; \i N
E
(P)
x
ij
0; 1; p
i
0 Z
; \(i; j) E; i N
E
:
The new formulation has several noticeable changes.
First, the new objective function has an extra term which
penalizes the pages that have excessive links, i.e., more
TABLE 3
The Set of All Candidate Links
TABLE 4
Relevant Candidate Links for b = 1
TABLE 2
An Example of Mini Sessions
than three links (C = 3) in this example. Specifically,
adding a new link into pages having three links or more
will result in a penalty which is five times (m = 5) as large
as the cost of adding a link into pages containing less than
three links. Second, constraint (P) is added to compute
exactly how many links exceed the out-degree threshold for
each page. Third, instead of having six constraints, the new
formulation has only four constraints corresponding to the
four relevant mini sessions. Two mini sessions (S
3
and S
6
)
no longer need to be facilitated when b increases to 2,
because they have only two paths and hence are not
considered in this example.
Increasing the path threshold also leads to more relevant
candidate links and subsequently a larger solution space for
each relevant mini session. As shown in Table 5, some
candidate links that were irrelevant for b = 1 will become
relevant to the decision when path threshold increases to
b = 2. For example, S
2
has two more relevant candidate
links for b = 2, i.e., (5, 6) and (1, 6), which can be used to
improve S
2
to achieve the user navigation goal for b = 2 but
not for b = 1. Note that S
5
strictly dominates S
2
, so if S
5
is
improved, S
2
will also be improved. Thus, the constraint
with respect to S
2
is redundant.
Solving the math program yields the optimal solution
x
26
= x
14
= x
56
= 1, with the other variables being 0. Since
all three links already exist according to the connectivity
matrix, no new link is needed at this time. It is worth
mentioning that when the multiplier for the penalty term is
set to m = 5 with the path threshold remaining at b = 1, the
optimal solution would change to x
16
= x
36
= x
64
= x
54
= 1.
Compared to the first case where m = 0 and b = 1, this case
adds two new links, i.e., (1, 6) and (3, 6), as opposed to only
one link, i.e., (4, 6). The reason is that after adding the
penalty term in the objective function, the MP model would
add (1, 6) and (3, 6) for a total cost of 2 (pages 1 and 3 have
only two links and are not penalized for having one more
link) instead of adding (4, 6) to incur a penalty of 5 (page 4
already has four links). Nevertheless, when the path
threshold increases to b = 2, no new link is needed because
fewer mini sessions are relevant to the decision and the
solution space for each relevant mini session increases.
5 COMPUTATIONAL EXPERIMENTS AND
PERFORMANCE EVALUATIONS
Extensive experiments were conducted, both on a data set
collected from a real website and on synthetic data sets. We
first tested the model with varying parameters values on all
data sets. Then, we partitioned the real data into training
and testing data. We used the training data to generate
improved site structures which were evaluated on the
testing data using two metrics that are discussed in detail
later. Moreover, we compared the results of our model with
that of a heuristic.
5.1 Real Data Set
5.1.1 Description of the Real Data Set
The real data set was collected from the Music Machines
website (http://machines.hyperreal.org) and contained
about four million requests that were recorded in a span of
four months. This data set is publicly available and has been
widely used in the literature [19], [29]. Table 6 shows the
number of pages in the website that had out-degrees within
a specified range. This website has in total 916 pages, of
which 716 have an out-degree of 20 or less, with the majority
(83 percent) of the remaining pages having 40 links or less.
Before analysis, we followed the log preprocessing steps
described in [29] to filter irrelevant information from raw
log files. These steps include: 1) filter out requests to pages
generated by Common Gateway Interface (CGI) or other
server-side scripts as we only consider static pages that are
designed as part of a website structure, 2) ignore un-
successful requests (returned HTTP status code not 200),
and 3) remove requests to image files (.gif, .jpg, etc.), as
images are in general automatically downloaded due to the
HTML tags rather than explicitly requested by users [33].
We utilized the page-stay time to identify target pages
and to demarcate mini sessions from the processed log files.
Three time thresholds (i.e., 1, 2, and 5 minutes) were used in
the tests to examine how results changes with respect to
different parameter values. Furthermore, we adapted the
algorithm proposed in [40] to identify the backtracking
pages in mini sessions, which are then used to demarcate
the paths traversed to reach the targets. Table 7 lists the
number of mini sessions comprising a given number of
paths (> 1) for different time thresholds.
5.1.2 Results and AnalysisReal Data Set
We set " = 1:0E 8 and vary the out-degree threshold (C),
the path threshold (b), andthe multiplier for the penalty term
(m) to examine how results change with respect to these
parameters. Table 8 reports the experiment results. The math
programs were coded in AMPL and solved using CPLEX/
AMPL 8.1 on a PC running Windows XP on an Intel Core 2
Duo E6300 processor. The times for generating optimal
TABLE 5
Relevant Candidate Links for b = 2
TABLE 6
Out-Degree Statistics
solutions varied from 0.109 to 0.938 seconds, indicating that
our model is very effective and practical for real-world
websites. We have reported the times taken to solve the math
programs only; the times taken for preprocessing steps and
obtaining values of a
S
ijkr
are not included, as they can be done
very quickly in practice. Note that the size of the real website
considered in our paper is significantly larger than the
average website size [52] as well as those used in related
papers addressing the website reorganization problem. For
example, Gupta et al. [19] and Lin and Tseng [28] report
results based on websites with only 427 and 146 pages,
respectively.
In Table 8, the column No. of new links indicates how
many new links need to be added into the current structure
in order to achieve the users navigation goal specified in the
column Path threshold (b) for all mini sessions. The
column No. of excessive links reports the number of new
links added to pages that have excessive links, i.e., have
more than C links. For example, if we set time threshold
to t = 2, out-degree threshold to C = 20, path threshold to
b = 1, and do not penalize pages for having excessive links
TABLE 7
Path Characteristics of Mini Sessions
TABLE 8
Results from the Real Data Set
(m = 0), then we need to establish 7,895 new links and
improve 1,711 existing links. Among all new links, 6,332 are
added to pages having 20 links or more. The numbers are
small considering the number of mini sessions needed for
improvement (30,292): approximately one new link is
needed per four mini sessions.
The number of new links needed decreases signifi-
cantly as the path threshold (b) increases. As discussed in
Section 4.2, a primary reason for this is that a large
number of mini sessions become irrelevant to the decision
as path threshold increases. In our test, when t = 2 and
b = 1 we have 30;292 mini sessions that are relevant and
will be considered in the MP model. This number
decreases to 6,807 as the path threshold increases to
b = 2: the other 23,485 ( = 30,292-6,807) mini sessions
containing two paths already meet the user navigation
goal and hence are no longer considered in the MP model
for b = 2.
The results show that when there is no penalty term
(m = 0), the number of new links and the number of
existing links to be improved are the same across different
out-degree thresholds (C). This is because the out-degree
threshold plays no role in the MP model if the penalty term
is removed from the objective function. When the penalty
term is imposed, i.e., m ,= 0, we find that while a larger
multiplier for the penalty term (m) leads to more new links,
it also adds fewer links to nodes having excessive links.
This is anticipated because as m increases, the MP model
would prefer to establish more links to pages with small
out-degrees in order to prevent large penalties.
Table 8 also indicates that, when the penalty term is
used, the use of a larger out-degree threshold (C) leads to
more links established. This occurs because, the larger the
out-degree threshold, the fewer the pages violating the out-
degree threshold. Thus, a larger out-degree threshold
provides the MP model with a larger space to add new
links without being penalized, and this results in more links
being established to pages having less than C links so that
heavy costs can be avoided when a penalty term is imposed
in the objective function.
5.1.3 Some Insights into Efficiency
The solution times taken to solve the math programs are
very low, consistently remaining below 1 second. Several
reasons contribute to the low solution times, as explained in
Section 4.2. The first reason is that a large number of
irrelevant mini sessions (those that need not to be improved
for a given path threshold) can be eliminated from
consideration. As a result, the number of relevant mini
sessions that will be improved and considered by the MP
model is not too large. Generally, increasing path thresholds
reduces the number of relevant mini sessions. Thus, the size
of relevant mini sessions is further reduced with increased
path thresholds. In our experiment, the number of relevant
mini sessions for t = 1 is 33,812 when path threshold is set
to b = 1, and this number decreases to 6,672 when the path
threshold increases to b = 2.
The second reason that can explain low solution times is
that our model formulation helps reduce the number of
relevant candidate links (the links can be selected to
improve user navigation to meet the goal). Particularly,
we identify the set of candidate links from mini sessions
instead of considering all possible links in a website, and
this significantly reduces the search space for our model.
Besides, as explained earlier, not all candidate links can help
relevant mini sessions meet the specified goal for user
navigation. Such candidate links are irrelevant to our
decision for a given path threshold and hence can be
eliminated from consideration. In the experiments, the total
number of candidate links for t = 1 is 23,032, and among
these links, the number of relevant candidate links for
b = 1; b = 2, and b = 3 are 18,373, 11,676, and 7,209, respec-
tively. Further, the number of relevant candidate links for
each relevant mini session was observed to be not very large
(generally consists of five links or less) and this leads to a
relatively small solution space for individual mini sessions.
The third reason for low solution times is that many mini
sessions dominate others with respect to relevant
candidate links. Consequently, the dominated mini sessions
need not to be considered in the constraints and can be
eliminated from our model. While the problem size can
be considerably reduced in our model formulation, the
remaining formulation continues to have a nontrivial
number of variables and constraints. We note that the data
set is from a real website and spans a four-month period,
which is long enough to cover most realistic situations.
5.2 Synthetic Data Sets
In addition to the real data set, synthetic/artificial data sets
were generated and considered for computational experi-
ments to evaluate the scalability of our model with respect
to the size of the website and the number of mini sessions.
For this reason, the artificial website structures and mini
sessions were generated to have similar statistical char-
acteristics as the real data set. For instance, the average out-
degree for pages in the real website is 15, so the link
structure for the artificial website was generated in a way
such that each page contained 15 links on average.
Three websites consisting of 1,000, 2,000, and 5,000 web-
pages were constructed. Our approach for generating the
link structure is similar to that described in [28]. Particu-
larly, to generate the link structure that contains 1,000 pages,
we first constructed a complete graph of 1,000 nodes
(pages) and each directed edge was assigned a random
value between 0 and 1. Then, we selected the edges with the
smallest 15,000 values to form the link structure, resulting in
an average out-degree of 15 for this website. In a similar
manner, we generated the link structures for other artificial
websites. The mini sessions were generated in a slightly
different manner. Specifically, we directly generated the set
of relevant candidate links for each mini session instead of
creating the users traversal path. As a result, this allows us
to directly apply the model on synthetic data sets. The sets
of relevant candidate links in synthetic data sets has similar
characteristics with those from the real one, comprising one
to five relevant candidate links per each relevant mini
session, with each link being randomly selected.
Each of the three artificial websites was tested with
10,000, 50,000, 100,000, and 300,000 mini sessions, resulting
in 12 different categories. In each category, three data sets
were generated and the results were averaged over the
three sets. We note that the synthetic data sets considered in
this paper are significantly larger than those used in related
papers, which report results based on synthetic data sets
with at most 200 pages and 3,200 links [28].
The math programs for the synthetic data were coded in
AMPL and solved using CPLEX/AMPL 11.1.1 on a PC
running Windows 7 on a 3.4 GHz processor. We
experimented the model with two out-degree thresholds,
i.e., C = 20 and C = 40, and two multipliers for the penalty
term, i.e., m = 0 and m = 5, on each synthetic data set.
Noticeably, the times for generating optimal solutions are
low for all cases and parameter values tested, ranging from
0.05 to 24.727 seconds. This indicates that the MP model is
very robust to a wide range of problem sizes and
parameter values. Particularly, the average solution times
for website with 1,000, 2,000, and 5,000 pages are 0.231,
1.352, and 3.148 seconds. While the solution times do go
up with the number of webpages, they seem to increase
within a reasonable range.
Besides these data sets, two large websites with 10,000
and 30,000 pages were generated and experimented with
300,000, 600,000, and 1.2 million mini sessions to emphasize
the fact that the model presented here is scalable to an even
larger extent. The solution times are also remarkably low
even in this case, varying from 1.734 to 33.967 seconds. In
particularly, the average solution times for websites with
10,000 and 30,000 pages are 3.727 and 6.086 seconds,
respectively. While the solution times also increase with the
size of the website, they seem to increase linearly or slower.
5.3 Evaluation of the Improved Website
In addition to the extensive computational experiments on
both real and synthetic data sets, we also perform
evaluations on the improved structure to assess whether
its navigation effectiveness is indeed enhanced by approx-
imating its real usage. Specifically, we partition the real data
set into a training set (first three months) and a testing set
(last month). We generate the improved structure using the
training data, and then evaluate it on the testing data using
two metrics: the average number of paths per mini session and
the percentage of mini sessions enhanced to a specified threshold.
The first metric measures whether the improved structure
can facilitate users to reach their targets faster than the
current one on average, and the second metric measures
how likely users suffering navigation difficulty can benefit
from the improvements made to the site structure. The
evaluation procedure using the first metric consists of three
steps and is described as follows:
1. Apply the MP model on the training data to obtain
the set of new links and links to be improved.
2. Acquire from the testing data the mini sessions that
can be improved, i.e., having two or more paths,
their length, i.e., number of paths, and the set of
candidate links that can be used to improve them.
3. For each mini session acquired in step 2, check
whether any candidate link matches one of the links
obtained in step 1, that is, the results from the
training data. If yes, with the assumption that users
will traverse the new link or the enhanced link in
the improved structure, remove all pages (excluding
the target page) visited after the source node of the
matching candidate link to obtain the new mini
session for the improved website, and get its updated
length information.
We illustrate the evaluation process with an example.
Let S
1
= 2; 1; 4; 5; 3; 6 and S
2
= 6; 3; 2; 5;
1; 4 be two mini sessions in testing data. Their current
lengths, i.e., their lengths in the current structure, are
L
m
(S
1
) = 4 and L
m
(S
2
) = 3. The set of candidate links for
S
1
is CL
1
= (2; 6); (1; 6); (4; 6); (5; 6) and for S
2
is CL
2
=
(6; 4); (3; 4); (2; 4); (5; 4). Let x
16
= x
46
= x
24
= 1 be the
results obtained from the MP model on training data.
Clearly, links (1, 6) and (4, 6) match two candidate links
for S
1
and link (2, 4) matches one candidate link for S
2
,
so we remove all nontarget pages after pages 1 and 4 in
S
1
and page 2 in S
2
. In this example, since page 1 is
visited prior to page 4 in S
1
, we remove all nontarget
pages after page 1 instead of page 4. The new mini
sessions for the improved structure are denoted by S
/
1
=
2; 1; 6 and S
/
2
= 6; 3; 2; 4 and their updated
lengths are L
m
(S
/
1
) = 1 and L
m
(S
/
2
) = 2, respectively.
The evaluation results using the first metric, the average
number of paths per mini session, are reported in Table 9. We
only report the results obtained from setting the threshold
t = 5, and the results obtained from other time thresholds
are similar. Note that the average of number of paths
needed to reach the targets before the improvements for
t = 5 is 2.383.
The results in Table 9 indicate that improved structures
can facilitate users to reach their targets in fewer paths than
TABLE 9
Evaluation Results on Improved Website Using Number of Paths Per Mini Session for T = 5 Min
the current one on average. As an example, when
m = 0; b = 2, and C = 40, the number of paths per mini
session in the improved structure decreases to 1.589 from
2.383, with the help of 1,145 new links. Note that while the
number of paths per mini session decreases with the
decreasing value of b for the MP model on the training data,
the number of new links that need to be established also
increases significantly. This is anticipated because, as
shown earlier, a smaller path threshold leads to more links
established, which results in fewer paths per mini session in
the improved structure as a consequence.
We now proceed to report the evaluation results using
the second metric, the percentage of mini sessions enhanced to a
specified threshold. We devise this metric to assess that, for
mini sessions whose current length are larger than a
specified evaluation threshold, how many of them are
enhanced such that their updated lengths in the improved
structure are no longer larger than the evaluation thresh-
old. In other words, we use this metric to measure how
many mini sessions, in which the current user navigation is
below the evaluation threshold, can benefit from the
improved structure.
We denote the evaluation threshold by b
e
where the
subscript e is used to distinguish the evaluation threshold
from the path threshold (b) that is used for the MP model on
the training set. The evaluation process is similar to the
steps for the first metric, but it needs one more step after
step 3, i.e., after length information is updated. Particularly,
for each mini session whose current length is larger than b
e
,
we check if its updated length decreases to b
e
or less; if yes,
then the mini session is considered enhanced by the
improved structure to satisfy the evaluation threshold.
We continue with the earlier example, where S
1
=
2; 1; 4; 5; 3; 6 and S
2
= 6; 3; 2; 5; 1; 4 are
in the testing set. When the evaluation threshold is set to
b
e
= 3, only S
1
needs to be evaluated, because the current
length of S
1
is larger than 3, i.e., L
m
(S
1
) = 4 > 3, whereas
the current length of S
2
, i.e., L
m
(S
2
) = 3, is not. Suppose the
solution from the training set is x
46
= x
64
= 1. We then
follow the three steps in the evaluation procedure and
update S
1
to S
/
1
= 2; 1; 4; 6. Since its updated length is
L
m
(S
/
1
) = 2 < b
e
= 3, S
1
is considered enhanced by the
improved structure. If the evaluation threshold changes to
b
e
= 1, then both S
1
and S
2
need to be evaluated, as their
current lengths are larger than 1. Similarly, we update S
1
and S
2
to S
/
1
= 2; 1; 4; 6 and S
/
2
= 6; 4. Now, only
S
2
is considered enhanced in the improved structure,
because its updated length is 1, while the length of S
/
1
is
still larger than 1. Three values of b
e
, i.e., b
e
= 1; 2, and 3,
were used for evaluation. We only report the results for
t = 5 in Table 10, and the results obtained using other time
thresholds are similar.
In Table 10, the values of b, m, and C are the parameters
used to generate the improved structures using training
data. The column Mini sessions enhanced (%) lists the
percentage of mini sessions that have more than b
e
paths in
the current structure but are enhanced to satisfy the
evaluation threshold, i.e., have b
e
or less paths, in the
improved structure. For example, when using m = 0;
C = 40, and b = 1 for the training data, 71.54 percent of
the mini sessions that have two or more paths in the current
structure are enhanced to have only one path (b
e
= 1) in the
improved structure. This means that the improved structure
helps more than 70 percent of users reach the targets in at
least one less path than the current structure.
As shown in Table 10, the use of a larger evaluation
threshold (b
e
) leads to more mini sessions enhanced. For
example, when using m = 1; C = 20, and b = 1 to generate
the improved structure, 70.80 and 87.57 percent of the mini
sessions are enhanced for b
e
= 1 and 3, respectively. In other
words, 87.57 percent of the users who traversed more than
three paths are facilitated in the improved structure
whereas only 70.80 percent of the users who traversed
more than one path are facilitated. This suggests that
severely disoriented users, i.e., those who need to exert
more effort to locate their targets, are more likely to benefit
from the improved site structure than the less disoriented
users, i.e., those who need to exert relatively less effort to
locate their targets. This is a very appealing result because
in practice the severely disoriented users are more likely to
abandon the website as compared to those who are not.
TABLE 10
Results of Evaluation on Improved Website Using Percentage of Mini Sessions Enhanced for T = 5 Min
We also observe that as the path threshold for the
training data (b) increases, more mini sessions are enhanced
in the testing data. This is because considerably more links
are added with a smaller value of b, as shown in Table 9.
However, adding more links may not necessarily always
lead to significantly superior results. For example, when
setting m = 0; C = 20, and b
e
= 1, while the use of b = 1
adds 5,794 new links, a number that is considerably larger
than the 1,145 new links added when setting b = 2, the
percentage of mini sessions enhanced increases only by
20.42 percent (= 71:54% 51:12%).
5.4 Comparison with a Heuristic
As mentioned earlier, although several transformation
approaches have been developed to restructure a website,
the objective of improving the user navigation on a website
with minimal changes to its structure has not yet been
examined. For this reason, we could not compare our model
with other approaches proposed in the previous studies. In
order to provide some comparisons, we compare the
performance of our model with that of a heuristic instead
and report the results in Table 11. The heuristic essentially
identifies the set of relevant candidate links for each mini
session and then randomly selects one link. We provide only
the comparison results for the case where m = 0 and C = 20
are set in our model, and the comparisons of the heuristic
with our model using other parameters are similar.
The first two columns of Table 11 list the parameters
used for both the heuristic and our model. The number of
new links and the number of improved links for the
heuristic are shown in the third and fourth columns. The
results of heuristic are obtained by performing the heuristic
three times and then taking the average. The five columns
under Difference = (Heuristic-Our model)/Our model
report the comparisons between the heuristic and our
model. Specifically, the first column lists how many more
new links are introduced by the heuristic than our model;
the next four columns show the increase/decrease in
performance by the heuristic as compared with our model.
For example, Table 11 shows that when the time threshold
and the path threshold are set to t = 1; b = 2 (second row),
the heuristic adds 112.12 percent more links than our
model. Meanwhile, it only reduces the average number of
paths per mini session by 11.42 percent and enhances
4.38 percent more mini sessions for b
e
= 2 than our model.
The comparison shows that, under the same path
threshold, the heuristic adds far more links than our
model, and the difference increases significantly as the
path threshold increases. For example, the heuristic adds
22.15 percent more new links for t = 1 and b = 1, but it
adds more than twice as many new links as our model
does (112.12 percent more links) when b increases to 2. The
result indicates that although the heuristic adds consider-
ably more links, it leads to only marginal improvements in
performance. Sometimes, our model outperforms the
heuristic with significantly fewer new links. For instance,
when t = 2; b = 3 (sixth row), while the heuristic enhances
3.99 percent more mini sessions when the evaluation
threshold is b
e
= 2, it adds 152.13 percent more new links
to the current structure than our model. Interestingly,
when b
e
= 3, our model even enhances more mini sessions
with much fewer new links. Overall, this comparison
shows that our model dominates this heuristic with respect
to the number of new links added while achieving
comparable or better results in facilitating user navigation.
6 DISCUSSION
6.1 Mini Session and Target Identification
We employed the page-stay timeout heuristic to identify
users targets and to demarcate mini sessions. The intuition
is that users spend more time on the target pages. Page-stay
time is a common implicit measurement found to be a good
indicator of page/document relevance to the user in a
number of studies [44], [53], [54]. In the context of web usage
mining, the page-stay timeout heuristic as well as other
time-oriented heuristics are widely used for session identi-
fication [29], [42], [43], [19], [55], and are shown to be quite
robust with respect to variations of the threshold values [56].
The identification of target pages and mini sessions can
be affected by the choice of page-stay timeout threshold.
Because it is generally very difficult to unerringly identify
TABLE 11
Comparison of Our Model with a Heuristic
mini sessions from anonymous user access data [19], we
ran our experiments for different threshold values.
Generally, increasing the threshold will result in fewer
mini sessions with proportionally more mini sessions
having a large number of paths while decreasing the
threshold will have the opposite effect (see Table 7). In
other words, increasing time threshold would decrease the
number of relevant mini session for small path thresholds
but could increase the number of relevant mini sessions
for large path thresholds. As a result, we observed that
as the time threshold increase, the number of new links
decreases for b = 1 and 2, but increases for b = 3. While
the results did change slightly, we showed that our model
succeeded in finding the minimal number of links that can
be used to improve user navigation substantially.
The time thresholds need to be properly selected based
on the amount of information displayed in the webpages. In
general, a larger time threshold is needed for websites
consisting of information-rich pages than those whose pages
contain less information. Other approaches can also help
accurately identify target pages. As an example, since web
designers have a good understanding of web contents, they
can identify a list of important pages with high probabilities
of being users targets. Such information could greatly help
improve the accuracy of target page identification.
6.2 Searching Sessions versus Browsing Sessions
While the majority of users typically have one or more goals
and search for particular pieces of information when
navigating a websites [57] (searching sessions), some
users might simply browse for general information
(browsing sessions). Though exact distinction between
these two web surfing behaviors is often impossible by only
looking at the anonymous user access data from weblogs,
certain characteristics can help differentiate the two types of
sessions. For example, some visits are clearly purposeless
and finish abruptly at pages that cannot be target pages, so
these sessions are more likely to be browsing sessions. To
the best of our knowledge, there is no algorithm developed
to distinguish between the two types of sessions and further
investigation on this question is needed. While we did not
explicitly separate searching sessions from browsing ses-
sions, the preprocessing steps can help eliminate many
purposeless browsing sessions. As a result, the final
improved website structure mainly shortens searching
sessions and also reduces purposeful browsing sessions.
6.3 Implications of This Research
This research contributes to the literature on improving
web user navigation by examining this issue from a new
and important angle. We have performed extensive experi-
ments on both real and synthetic data sets to show that the
model can be effectively solved and is highly scalable. In
addition, the evaluation results confirm that users can
indeed benefit from the improved structure after suggested
changes are applied. There are several important implica-
tions from this research.
First, we demonstrate that it is possible to improve user
navigation significantly with only few changes to the
structure of a website using the proposed model. This is
important because as time passes and the need for
information changes, websites also need to be regularly
maintained and improved. However, the current literature
focuses on how to restructure a website and hence is not
suitable for this purpose. Our research complements the
literature by addressing this issue using a MP model. While
the approaches proposed in previous studies are either
heuristic-based or applicable only to small-sized website, we
have shown that our model not only can be effectively
solved for optimal solutions but also can scale up well, partly
due to the fact that our model formulation can reduce the
problem size considerably.
Second, we model the out-degree as a penalty (cost) term
in the objective function, and this not only leads to more
flexible website structures than modeling the out-degree as
hard constraints, but also offers a good balance between
minimizing changes to the website and reducing informa-
tion overload to users (avoid clustering too many links in
webpages). In particular, the experiment results indicate
that when the penalty term is used, although more links are
needed, the number of links inserted into pages with large
out-degrees is also significantly reduced. This helps prevent
further adding links to page with many links and helps the
users locate the desired links in these pages more easily.
Third, we show that the use of a small path threshold will
introduce far more changes to a website but may not always
lead to significantly better outcomes for user navigation.
This suggests that Webmasters need to carefully balance the
tradeoff between desired improvements in the user naviga-
tion and the changes needed to accomplish the task when
selecting appropriate path thresholds. This is particularly
important when a website is improved on a regular basis.
Last, the proposed model can be exceptionally desirable
to severely disoriented users. The evaluation results show
that severely disoriented users, i.e., those who need to exert
more effort to locate the targets, are more likely to benefit
from the changes suggested by our model than the less
disoriented users. This is an appealing result because the
severely disoriented users are more likely to abandon their
search for targets as compared to those who are not.
6.4 Choice of Parameter Values for the Model
6.4.1 Path Threshold
The path threshold represents the goal for user navigation
that the improved structure should meet and can be
obtained in several ways. First, it is possible to identify
when visitors exit a website before reaching the targets from
analysis of weblog files [40], [58]. Hence, examination of
these sessions helps make a good estimation for the path
thresholds. Second, surveying website visitors can help
better understand users expectations and make reasonable
selections on the path threshold values. For example, if the
majority of the surveyed visitors respond that they usually
give up after traversing four paths, then the path threshold
should be set to four or less. Third, firms like comScore and
Nielsen have collected large amounts of client-side web
usage data over a wide range of websites. Analyzing such
data sets can also provide good insights into the selection of
path threshold values for different types of websites.
Although using small path thresholds could result in
more improvements in web user navigation in general, our
experiments showed that the changes (costs) needed
increase significantly as the path threshold decreases.
Sometimes, additional improvements in user navigation
from using a small threshold are too little to justify the
increased costs. Thus, Webmasters need to cautiously
consider the tradeoff between desired improvements to
user navigation and the changes needed when selecting
appropriate values for path threshold. A cost benefit
analysis that compares benefits and costs of using
different path thresholds can be useful for this purpose. In
the context of our problem, we can view the number of new
links needed as the cost and the improvement on user
navigation (this, for instance, can be measured as the
average number of paths shortened by the improved
structure) as the benefit. The benefit-cost ratio (BCR) that
is used for the analysis of the cost effectiveness of different
options can be expressed as (improvement on user
navigation)/(number of new links). As a result, we can
employ the evaluation procedures described in Section 5.3
to approximate the BCRs for different path thresholds and
to find the path threshold value that is the most cost
effective and provides a good tradeoff. In addition, our
experiments showed that an extremely small path threshold
(e.g., b = 1) could add many links and hence might not be a
good choice if Webmasters plan to use our model to
improve websites progressively.
6.4.2 Out-Degree Threshold
Webpages can be generally classified into two categories
[29]: index pages and content pages. An index page is
designed to help users better navigate and could include
many links, while a content page contains information users
are interested in and should not have many links. Thus, the
out-degree threshold for a page is highly dependent on the
purpose of the page and the website. Typically, the out-
degree threshold for index pages should be larger than that
for content pages. For instance, out-degree thresholds are
set to 30 and 10 for index and content pages, respectively, in
the experiments in [29]. Since out-degree thresholds are
context dependent and organization dependent, behavioral
and experimental studies that examine the optimal out-
degree threshold for different settings are needed. In
general, the out-degree threshold could be set at a small
value when most webpages have relatively few links, and
as new links are added, the threshold can be gradually
increased. Note that since our model does not impose hard
constraints on the out-degrees for pages in the improved
structure, it is less affected by the choices of out-degree
thresholds as compared to those in the literature.
6.4.3 Multiplier for the Penalty Term
As shown in Section 5, the use of the penalty term can
prevent the model from adding new links to pages that
already have many links. This helps keep the information
load low for user at the cost of inserting more new links into
other pages with small out-degrees. Generally, if a website
have both pages with small out-degrees and pages with
very large out-degrees, then it is reasonable to use a large
multiplier (m) to avoid clustering too many links in a page.
If the out-degrees are relatively small for all pages, then it
could be more appropriate to use a relatively small multi-
plier to minimize the total number of new links added.
When our model is used for website maintenance, a small
multiplier could be used in the beginning when out-degrees
are generally small for most pages, and as new links are
inserted, a larger multiplier is needed to prevent adding
extra links to pages that already have many links. In general,
we suggest not using a very large multiplier for the penalty
term, because the benefit from the reduced information load
by using the penalty term was observed to improve only
very slightly as the multiplier increases in the experiments.
In particular, as shown in Table 8, increasing m from 0 to 1
would prevent far more links from adding to pages with
large out-degrees (i.e., reduce far more excessive links)
than increasing m from 1 to 5, while the numbers of new
links needed were similar in the two cases.
6.5 Evaluation Procedure
We used simulations to approximate the real usage and to
evaluate how the user navigation could be enhanced in the
improved website structure. The use of simulation for
website usability evaluation is very popular and has been
widely used in modeling users choices in web navigation
and usability test [34], [36], [37], [59], [60]. However,
simulation studies often have to make simplifying assump-
tions in order to simulate real-life scenarios, posing
questions on the generalizability of the results. In the
context of our simulation approach, we assume that users
would find their target pages effectively through a new/
improved link if it exists. In practice, certain criteria related
to the visual design of web interfaces need to be followed in
order to effectively apply the suggested changes to a
website. We note that there exist an abundant literature
on both webpage design [6], [8], [12], [31], [59], [61] and
hyperlink design [34], [49], [50], [62], [63], [64]. Though we
did not explicitly consider design issues in this paper, we
do assume that Webmasters follow the guidelines and
suggestions from such studies when creating and editing
links and designing webpages. Consequently, in the
simulation approach used for user navigation evaluation,
we assume that new links are carefully designed and
existing links are appropriately edited. In addition, they
should also be placed in proper places for users to easily
locate. Thus, these links should provide users with accurate
knowledge on the contents on the other end of a link and
help them make correct selections.
Because of the assumption made for the new and
improved links, the claimed benefit can be interpreted as
the upper bound and optimal benefit of our model.
However, we would like to claim that improved and newly
added links could guide users to find their target pages
more efficiently to some extent. This is because: 1) our
method establishes efficient paths to target pages that were
not available in the website structure before optimization,
and 2) our method suggests improving links that would
lead to users target pages efficiently but missed by users
(since they did not know what these links would lead to), so
that more efficient navigation can be facilitated.
Since our evaluation is simulation based, a usability
study involving real users may help strengthen the results
of our study and deserves further investigation. However,
we note that such usability studies are generally more
expensive and time consuming in the context of website
evaluation [60], and hence are usually conducted on small-
sized websites [5]. In contrast, simulation can be easily
implemented, quickly performed for various parameter
settings, and tested on a large scale. Thus, the simulation
studies in our paper complement usability studies by
offering its own distinct advantages.
7 CONCLUSIONS
In this paper, we have proposed a mathematical program-
ming model to improve the navigation effectiveness of a
website while minimizing changes to its current structure, a
critical issue that has not been examined in the literature.
Our model is particularly appropriate for informational
websites whose contents are relatively stable over time. It
improves a website rather than reorganizes it and hence is
suitable for website maintenance on a progressive basis.
The tests on a real website showed that our model could
provide significant improvements to user navigation by
adding only few new links. Optimal solutions were quickly
obtained, suggesting that the model is very effective to real-
world websites. In addition, we have tested the MP model
with a number of synthetic data sets that are much larger
than the largest data set considered in related studies as
well as the real data set. The MP model was observed to
scale up very well, optimally solving large-sized problems
in a few seconds in most cases on a desktop PC.
To validate the performance of our model, we have
defined two metrics and used them to evaluate the
improved website using simulations. Our results con-
firmed that the improved structures indeed greatly
facilitated user navigation. In addition, we found an
appealing result that heavily disoriented users, i.e., those
with a higher probability to abandon the website, are more
likely to benefit from the improved structure than the less
disoriented users. Experiment results also revealed that
while using small path thresholds could result in better
outcomes, it would also add significantly more new links.
Thus, Webmasters need to carefully balance the tradeoff
between desired improvements to the user navigation and
the number of new links needed to accomplish the task
when selecting appropriate path thresholds. Since no prior
study has examined the same objective as ours, we
compared our model with a heuristic instead. The
comparison showed that our model could achieve compar-
able or better improvements than the heuristic with
considerably fewer new links.
The paper can be extended in several directions in
addition to those mentioned in Section 6. For example,
techniques that can accurately identify users targets are
critical to our model and future studies may focus on
developing such techniques. As another example, our
model has a constraint for out-degree threshold, which is
motivated by cognitive reasons. The model could be further
improved by incorporating additional constraints that can
be identified using data mining methods [65]. For instance,
if data mining methods find that most users access the
finance and sports pages together, then this information can
be used to construct an additional constraint.
ACKNOWLEDGMENTS
The authors would like to thank the editors and the
anonymous reviewers for their insightful comments and
helpful suggestions, which have resulted in substantial
improvements to this work.
REFERENCES
[1] Pingdom, Internet 2009 in Numbers, http://royal.pingdom.
com/2010/01/22/internet-2009-in-numbers/, 2010.
[2] J. Grau, US Retail e-Commerce: Slower But Still Steady Growth,
http://www.emarketer.com/Report.aspx?code=emarketer_
2000492, 2008.
[3] Internetretailer, Web Tech Spending Static-But High-for the
Busiest E-Commerce Sites, http://www.internetretailer.com/
dailyNews.asp?id = 23440, 2007.
[4] D. Dhyani, W.K. Ng, and S.S. Bhowmick, A Survey of Web
Metrics, ACM Computing Surveys, vol. 34, no. 4, pp. 469-503, 2002.
[5] X. Fang and C. Holsapple, An Empirical Study of Web Site
Navigation Structures Impacts on Web Site Usability, Decision
Support Systems, vol. 43, no. 2, pp. 476-491, 2007.
[6] J. Lazar, Web Usability: A User-Centered Design Approach. Addison
Wesley, 2006.
[7] D.F. Galletta, R. Henry, S. McCoy, and P. Polak, When the Wait
Isnt So Bad: The Interacting Effects of Website Delay, Familiarity,
and Breadth, Information Systems Research, vol. 17, no. 1, pp. 20-
37, 2006.
[8] J. Palmer, Web Site Usability, Design, and Performance Metrics,
Information Systems Research, vol. 13, no. 2, pp. 151-167, 2002.
[9] V. McKinney, K. Yoon, and F. Zahedi, The Measurement of Web-
Customer Satisfaction: An Expectation and Disconfirmation
Approach, Information Systems Research, vol. 13, no. 3, pp. 296-
315, 2002.
[10] T. Nakayama, H. Kato, and Y. Yamane, Discovering the Gap
between Web Site Designers Expectations and Users Behavior,
Computer Networks, vol. 33, pp. 811-822, 2000.
[11] M. Perkowitz and O. Etzioni, Towards Adaptive Web Sites:
Conceptual Framework and Case Study, Artificial Intelligence,
vol. 118, pp. 245-275, 2000.
[12] J. Lazar, User-Centered Web Development. Jones and Bartlett
Publishers, 2001.
[13] Y. Yang, Y. Cao, Z. Nie, J. Zhou, and J. Wen, Closing the Loop in
Webpage Understanding, IEEE Trans. Knowledge and Data Eng.,
vol. 22, no. 5, pp. 639-650, May 2010.
[14] J. Hou and Y. Zhang, Effectively Finding Relevant Web Pages
from Linkage Information, IEEE Trans. Knowledge and Data Eng.,
vol. 15, no. 4, pp. 940-951, July/Aug. 2003.
[15] H. Kao, J. Ho, and M. Chen, WISDOM: Web Intrapage
Informative Structure Mining Based on Document Object Model,
IEEE Trans. Knowledge and Data Eng., vol. 17, no. 5, pp. 614-627,
May 2005.
[16] H. Kao, S. Lin, J. Ho, and M. Chen, Mining Web Informative
Structures and Contents Based on Entropy Analysis, IEEE Trans.
Knowledge and Data Eng., vol. 16, no. 1, pp. 41-55, Jan. 2004.
[17] C. Kim and K. Shim, TEXT: Automatic Template Extraction from
Heterogeneous Web Pages, IEEE Trans. Knowledge and Data Eng.,
vol. 23, no. 4, pp. 612-626, Apr. 2011.
[18] M. Kilfoil et al., Toward an Adaptive Web: The State of the Art
and Science, Proc. Comm. Network and Services Research Conf.,
pp. 119-130, 2003.
[19] R. Gupta, A. Bagchi, and S. Sarkar, Improving Linkage of Web
Pages, INFORMS J. Computing, vol. 19, no. 1, pp. 127-136, 2007.
[20] C.C. Lin, Optimal Web Site Reorganization Considering In-
formation Overload and Search Depth, European J. Operational
Research, vol. 173, no. 3, pp. 839-848, 2006.
[21] M. Eirinaki and M. Vazirgiannis, Web Mining for Web
Personalization, ACM Trans. Internet Technology, vol. 3, no. 1,
pp. 1-27, 2003.
[22] B. Mobasher, H. Dai, T. Luo, and M. Nakagawa, Discovery and
Evaluation of Aggregate Usage Profiles for Web Personalization,
Data Mining and Knowledge Discovery, vol. 6, no. 1, pp. 61-82, 2002.
[23] B. Mobasher, R. Cooley, and J. Srivastava, Automatic Persona-
lization Based on Web Usage Mining, Comm. ACM, vol. 43, no. 8,
pp. 142-151, 2000.
[24] B. Mobasher, R. Cooley, and J. Srivastava, Creating Adaptive
Web Sites through Usage-Based Clustering of URLs, Proc.
Workshop Knowledge and Data Eng. Exchange, 1999.
[25] W. Yan, M. Jacobsen, H. Garcia-Molina, and U. Dayal, From User
Access Patterns to Dynamic Hypertext Linking, Computer Net-
works and ISDN Systems, vol. 28, nos. 7-11, pp. 1007-1014, May 1996.
[26] M. Nakagawa and B. Mobasher, A Hybrid Web Personalization
Model Based on Site Connectivity, Proc. Web Knowledge Discovery
Data Mining Workshop, pp. 59-70, 2003.
[27] B. Mobasher, Data Mining for Personalization, The Adaptive
Web: Methods and Strategies of Web Personalization, A. Kobsa, W.
Nejdl, P. Brusilovsky, eds., vol. 4321, pp. 90-135, Springer-Verlag,
2007.
[28] C.C. Lin and L. Tseng, Website Reorganization Using an Ant
Colony System, Expert Systems with Applications, vol. 37, no. 12,
pp. 7598-7605, 2010.
[29] Y. Fu, M.Y. Shih, M. Creado, and C. Ju, Reorganizing Web Sites
Based on User Access Patterns, Intelligent Systems in Accounting,
Finance and Management, vol. 11, no. 1, pp. 39-53, 2002.
[30] M.D. Marsico and S. Levialdi, Evaluating Web Sites: Exploiting
Users Expectations, Intl J. Human-Computer Studies, vol. 60, no. 3,
pp. 381-416, 2004.
[31] J. Palmer, Designing for Web Site Usability, Computer, vol. 35,
no. 7, pp. 102-103, June 2002.
[32] J. Liu, S. Zhang, and J. Yang, Characterizing Web Usage
Regularities with Information Foraging Agents, IEEE Trans.
Knowledge and Data Eng., vol. 16, no. 5, pp. 566-584, May 2004.
[33] P. Pirolli and S.K. Card, Information Foraging, Psychological
Rev., vol. 106, no. 4, pp. 643-675, 1999.
[34] E.H. Chi, P. Pirolli, and J. Pitkow, The Scent of a Site: A System
for Analyzing and Predicting Information Scent, Usage, and
Usability of a Web Site, Proc. SIGCHI Conf. Human Factors in
Computing Systems, pp. 161-168, 2000.
[35] C. Olston and E.H. Chi, ScentTrails: Integrating Browsing and
Searching on the Web, ACM Trans. Computer-Human Interaction,
vol. 10, no. 3, pp. 177-197, 2003.
[36] E.H. Chi, P. Pirolli, K. Chen, and J. Pitkow, Using Information
Scent to Model User Information Needs and Actions on the Web,
Proc. ACM Conf. Human Factors in Computing Systems, pp. 490-497,
2001.
[37] W.T. Fu and P. Pirolli, SNIF-ACT: A Cognitive Model of User
Navigation on the World Wide Web, Human-Computer Interac-
tion, vol. 22, pp. 355-412, 2007.
[38] W. Willett, J. Heer, and M. Agrawala, Scented Widgets:
Improving Navigation Cues with Embedded Visualizations,
IEEE Trans. Visualization and Computer Graphics, vol. 13, no. 6,
pp. 1129-1136, Nov. 2007.
[39] X. Xie, H. Liu, W. Ma, and H. Zhang, Browsing Large Pictures
under Limited Display Sizes, IEEE Trans. Multimedia, vol. 8, no. 4,
pp. 707-715, Aug. 2006.
[40] R. Srikant and Y. Yang, Mining Web Logs to Improve Web Site
Organization, Proc. 10th Intl Conf. World Wide Web, pp. 430-437,
2001.
[41] M.S. Chen, J.S. Park, and P.S. Yu, Efficient Data Mining for Path
Traversal Patterns, IEEE Trans. Knowledge and Data Eng., vol. 10,
no. 2, pp. 209-221, Mar./Apr. 1998.
[42] R. Cooley, B. Mobasher, and J. Srivastava, Data Preparation for
Mining World Wide Web Browsing Patterns, Knowledge and
Information Systems, vol. 1, pp. 1-27, 1999.
[43] M. Spiliopoulou, B. Mobasher, B. Berendt, and M. Nakagawa, A
Framework for the Evaluation of Session Reconstruction Heur-
istics in Web-Usage Analysis, INFORMS J. Computing, vol. 15,
no. 2, pp. 171-190, 2003.
[44] M. Morita and Y. Shinoda, Information Filtering Based on User
Behavior Analysis and Best Match Text Retrieval, Proc. 17th Ann.
Intl ACMSIGIR Conf. Research and Development in Information
Retrieval, pp. 272-281, 1994.
[45] Tealeaf, The Two Waves of Online Abandonment: The 2007
Harris Interactive Survey of Online Customer Behavior, http://
www.tealeaf.com/downloads/tealeaf-executivebrief_Harris2007.
pdf, 2007.
[46] J. Song and F.M. Zahedi, A Theoretical Approach to Web Design
in E-Commerce: A Belief Reinforcement Model, Management
Science, vol. 51, no. 8, pp. 1219-1235, 2006.
[47] V. Venkatesh and R. Agarwal, From Visitors into Customers: A
Usability-Centric Perspective on Purchase Behavior in Electronic
Channels, Management Science, vol. 52, no. 3, pp. 367-382, 2006.
[48] W. Lin, S. Alvarez, and C. Ruiz, Efficient Adaptive-Support
Association Rule Mining for Recommender Systems, Data Mining
and Knowledge Discovery, vol. 6, pp. 83-105, 2002.
[49] K. Larson and M. Czerwinski, Web Page Design: Implications of
Memory, Structure and Scent for Information Retrieval, Proc.
SIGCHI Conf. Human Factors in Computing Systems, pp. 25-32, 1998.
[50] M. Otter and H. Johnson, Lost in Hyperspace: Metrics and
Mental Models, Interacting with Computers, vol. 13, pp. 1-40, 2000.
[51] M. Garey and D. Johnson, Computers and Intractability: A Guide to
the Theory of NP-Completeness. W.H. Freeman, 1979.
[52] T. Boutell, WWW FAQs: How Many Websites Are There?
http://www.boutell.com/newfaq/misc/sizeofweb.html, 2007.
[53] D.W. Oard and J. Kim, Modeling Information Content Using
Observable Behavior, Proc. ASIST Ann. Meeting, pp. 481-488,
2001.
[54] M. Claypool, P. Le, M. Waseda, and D. Brown, Implicit Interest
Indicators, Proc. Sixth Intl Conf. Intelligent User Interfaces, pp. 33-
40, 2001.
[55] H. Liu and V. Keselj, Combined Mining of Web Server Logs and
Web Contents for Classifying User Navigation Patterns and
Predicting Users Future Requests, Data and Knowledge Eng.,
vol. 61, no. 2, pp. 304-330, 2007.
[56] B. Berendt, B. Mobasher, M. Spiliopoulou, and J. Wiltshire,
Measuring the Accuracy of Sessionizers for Web Usage Analy-
sis, Proc. Web Mining Workshop First SIAM Intl Conf. Data Mining,
pp. 7-14, 2001.
[57] J. Morrison, P. Pirolli, and S.K. Card, A Taxonomic Analysis of
What World Wide Web Activities Significantly Impact Peoples
Decisions and Actions, Proc. ACM Conf. Human Factors in
[58] R.E. Bucklin and C. Sismeir, A Model of Website Browsing
Behavior Estimated on Clickstream Data, J. Marketing Research,
vol. 40, no. 3, pp. 249-267, 2003.
[59] M. Ivory, R.R. Sinha, and M. Hearst, Empirically Validated Web
Page Design Metrics, Proc. SIGCHI Conf. Human Factors in
[60] C.S. Miller and R.W. Remington, Modeling Information Naviga-
tion: Implications for Information Architecture, Human Computer
Interaction, vol. 19, pp. 225-271, 2004.
[61] M. Ivory and M. Hearst, Improving Web Site Design, IEEE
Internet Computing, vol. 6, no. 2, pp. 56-63, Mar. 2002.
[62] J. Nielsen, Designing Web Usability: The Practice of Simplicity. New
Riders Publishing, 2000.
[63] J.M. Spool, T. Scanlon, W. Schroeder, C. Snyder, and T. DeAngelo,
Web Site Usability: A Designers Guide. Morgan Kaufman, 1999.
[64] J. Kim and B. Yoo, Toward the Optimal Link Structure of the
Cyber Shopping Mall, Intl J. Human-Computer Studies, vol. 52,
no. 3, pp. 531-551, 2000.
[65] B. Padmanabhan and A. Tuzhilin, On the Use of Optimization
for Data Mining: Theoretical Interactions and eCRM Opportu-
nities, Management Science, vol. 49, no. 10, pp. 1327-1343, 2003.
Min Chen received the PhD degree in manage-
ment science from the School of Management,
The University of Texas at Dallas, in 2011. He is
an assistant professor in the School of Manage-
ment at George Mason University. His research
interests include data mining, machine learning,
optimization models, and economics of informa-
tion systems.
Young U. Ryu received the PhD degree in
management science and information systems
from the McCombs Graduate School of Busi-
ness, The University of Texas at Austin. He is an
associate professor of information systems in
the School of Management, The University of
Texas at Dallas. He has studied applications of
data mining and artificial intelligence technolo-
gies and decision science methods to the
modeling and analysis of information systems.
He currently works on data separation as information classification and
machine learning, computer security, and social network analysis.
. For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.

LSJ1309 - Facilitating Effective User Navigation Through Website Structure Improvement

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

LSJ1309 - Facilitating Effective User Navigation Through Website Structure Improvement

Uploaded by

Copyright:

Available Formats

Facilitating Effective User Navigation through

Website Structure Improvement

You might also like